Can AI Correct Health Misinformation? Evidence from Correction Source and Message Strategies

Health misinformation poses a persistent challenge to public understanding and health-related decision-making. With the increasing availability and use of artificial intelligence (AI) systems for health information seeking, AI has begun to function not only as an information provider but also as a potential correction source for addressing health-related misconceptions. However, little is known about how AI compares with traditional institutional sources, such as public health organizations, in correcting health misinformation, nor how message strategies may shape the effectiveness of such corrections. This study investigates how different correction sources and message strategies influence individuals’ endorsement of a health-related misinformation claim. Using a between-subjects online experiment, participants are exposed to a corrective message addressing a commonly circulating health misconception related to pregnancy and seasonal influenza vaccination. The experiment adopts a 3 (correction source: AI health assistant vs. public health institution vs. text-only control) × 2 (message strategy: one-sided vs. two-sided) factorial design. Correction source is operationalized through visual presentation formats that attribute the same corrective content to different sources (e.g., an AI chat interface or a public health information page), while message strategy varies whether the corrective message solely presents the corrective perspective (one-sided) or explicitly acknowledges and refutes a common concern before presenting the correction (refutational two-sided). The substantive content of the corrective message is held constant across conditions to isolate the effects of source attribution and message strategy. The primary outcome is participants’ endorsement of the misinformation claim, measured as perceived correctness of the claim after exposure to the corrective message. Secondary outcomes include trust in the correction source, trust in the information, psychological reactance, perceived appropriateness of AI for health misinformation correction, and intentions to share the information. Baseline belief may be measured prior to message exposure to allow for robustness checks and exploratory analyses, but the primary analyses focus on post-correction belief endorsement under randomized assignment. By examining AI as a correction source alongside institutional and non-attributed sources, this study aims to clarify whether and how AI-mediated correction differs from traditional public health communication, and how message design shapes the effectiveness of misinformation correction in contemporary health information environments.